Smartphone Edge-Analytics for Real-Time Food Quality Assessment Using Spectroscopic Sensors

Thesis Abstract

This study addresses the growing need for rapid, low-cost, and scalable food quality assessment by leveraging smartphone edge analytics integrated with spectroscopic sensors to deliver real-time, non-destructive evaluations across diverse supply chains. The central aim is to develop and validate an edge-enabled framework that processes spectroscopic data on mobile devices to predict key quality attributes—such as ripeness, contamination, and freshness—of fruits, vegetables, and dairy products with high accuracy. Specific objectives include (1) to design and implement a smartphone-based data acquisition pipeline combining low-cost spectrometers (visible-NIR range) with edge computing modules; (2) to develop robust machine learning models (partial least squares regression, support vector regression, and lightweight neural networks) for quantitative quality prediction; (3) to evaluate model generalizability across species, processing stages, and storage conditions; (4) to assess real-time performance metrics, including prediction latency, energy consumption, and on-device memory usage; (5) to validate the approach against laboratory reference methods (gas chromatography–mass spectrometry for volatile compounds, high-performance liquid chromatography for biochemical markers) and standard sensory panels; and (6) to formulate deployment guidelines and ethical considerations for field use. Methodologically, a mixed-methods approach is adopted within a pragmatic research design. The population comprises packaged and unpackaged produce and dairy items sourced from three metropolitan markets and two regional distribution centers. A sample of 600 items is stratified by product type (15 categories), harvest stage, and prior storage duration. Instrumentation includes smartphone platforms running a custom edge analytics app, a compact spectrometer (280–1100 nm), calibrated reference spectrophotometers in the lab for ground-truth data, and portable analytical instruments for validation (LC-MS for metabolites, GC-MS for volatiles). Data collection combines (i) spectral data collected in controlled and field conditions (n ? 18,000 spectra), (ii) corresponding reference quality measurements including chemical markers and calibrated sensory scores, and (iii) device performance logs (latency, battery consumption, and CPU utilization). Instrument validity is ensured via cross-calibration against laboratory standards and repeat measurements (2–3 replicates per item). Data analysis proceeds in three stages. First, spectral preprocessing includes baseline correction, normalization, and feature extraction via principal component analysis and wavelength-wise variable importance in projection (VIP) scores. Second, predictive modeling employs an ensemble of models—partial least squares regression (PLSR), ridge regression, support vector regression with a radial basis function kernel, and a compact convolutional neural network tailored for on-device inference—implemented with model compression techniques (quantization and pruning) to meet device constraints. Third, model evaluation uses nested cross-validation to assess predictive accuracy (root mean square error, R^2), robustness across product categories, and transferability to unseen storage conditions. Hypothesis testing includes ANOVA to compare model performances across product groups and regression diagnostics to assess multicollinearity and heteroscedasticity. The study also examines theoretical underpinnings from the Theory of Technological Frames and Diffusion of Innovations to interpret user acceptance and deployment implications. Expected findings indicate that edge-processed spectroscopic data can predict key quality indicators with R^2 values above 0.85 for several attributes (e.g., freshness scores, moisture content, and volatile markers) and with mean absolute errors within industry-acceptable ranges. The on-device models are anticipated to achieve prediction latencies under 2 seconds per item and energy consumption compatible with routine smartphone use in field settings. It is expected that model generalization will be strongest within product categories with similar spectral signatures and weaker across highly heterogeneous items, necessitating periodic model recalibration. The study contributes to knowledge by demonstrating the viability of fully on-device spectroscopy for real-time food quality assessment, bridging spectroscopy, edge computing, and mobile analytics, and by offering a deployable framework with guidelines for data governance, privacy, and user-centric design. Conclusions anticipate that smartphone edge analytics can reliably augment food quality decision-making in supply chains, enabling rapid, non-destructive screening and reducing waste through timely interventions. Recommendations include (1) expanding reference libraries to improve cross-category generalization; (2) integrating cloud-backed model update mechanisms for continual learning; (3) developing standardized protocols for field calibration and data privacy; and (4) pursuing collaborations with industry partners to pilot large-scale deployments in retail and logistics environments.